IVS Briefs
IVS / Edge AI Research Brief
Signals (8)
- 1
Key point
Board of Directors 5–4 Tuesday night terminated Manchester’s Flock contract and ordered all 18 ALPRs shut down, covered, then removed before year-end. Hours of comment; most speakers opposed. Fourth Connecticut town to end Flock, after Windsor, Killingworth, and Ledyard. Supporters cited a Sunday stabbing arrest.
Why it matters
A cover-then-remove by Dec 31 unwind, not a 2027 expire. Bid a local VMS/LPR only if the Board wants a replacement; the 5–4 is a take-down, not a vendor swap.
Source
NBC Connecticut - 2
Key point
City Council 3–2 Tuesday (Mayor Michael Vargas and David Starr Rabb no) suspended a two-year $615,000 Flock agreement. Riverside County Sheriff’s Perris Station: 104 ALPRs. Original March 2024 install was ~$650,000; that term expired June. Marisela Nava and Malcolm Corona backed the suspend until safeguards exist; Corona: Flock “has lost the public’s trust.” RSO Capt. Kevin Lamb and Lt. Deirdre Vickers: losing plates delays stolen-vehicle and missing-person work. KVCR: first Inland Empire city to suspend Flock (Coachella rejected ~70 cams in 2023). No take-down date in the KVCR write-up.
Why it matters
A 104-cam / $615k suspend, not a dark-and-done. Quote suspend-until-safeguards, not terminate, and do not certify bags from Tuesday’s 3–2.
Source
KVCR - 3
Key point
Town Council Tuesday approved a resolution adding oversight and reporting for the town’s Flock cameras and other ALPRs. WCBD: months of privacy / retain / use debate; leaders framed it as transparency, not a ban. Cameras stay.
Why it matters
A keep-with-rules night, not a bag night. Any Flock add or swap in Mount Pleasant now runs through the new reporting lock.
Source
CountOn2 / WCBD - 4
Key point
City Manager James Parajon told council the evaluation resolution for any future surveillance tech needs “a little bit more time” and will come Nov 24. He had promised a Nov 10 return on Oct 27; the item was off the Nov 6 docket. Modeled on the ACLU template used in other cities; staff is working with ACLU, council, and residents. No draft released. Hardware decommission was already numbered 10-30; this is the replacement rubric, still unvoted.
Why it matters
Do not bid a Flock/Axon/Motorola refill into Alexandria before Nov 24. The hole is real; the rules are not.
Source
The Alexandria Brief - 5
PureTech PureActiv → Genetec Security Center 5.13.2.0: live bounding boxes, alarms, health in the existing VMS
Key point
PureTech Systems posted an enhanced PureActiv integration with Genetec Security Center 5.13.2.0: classification bounding boxes on live video, PureActiv alarms (plus video-loss / lost-comms / analytics-hindered health) inside Genetec alarm handling. Metadata includes alarm description and associated camera. CEO Larry Bowe: no new console, no new training. Aimed at perimeter with low nuisance-alarm detection.
Why it matters
A Genetec-native perimeter-analytics attach for sites that already standardized on Security Center. Quote 5.13.2.0 as the floor and demo boxes-on-live plus health alarms, not a side GUI.
- 6
HiPFuse: real-time LiDAR semantic segmentation on Jetson AGX Orin — 14.2 / 22.4 FPS, beats 10 Hz scan rate
Key point
Haiyang Wu et al.: HiPFuse does LiDAR semantic segmentation on edge GPUs with height-preserving BEV fusion and dual-scale context. SemanticKITTI 63.7% mIoU at 14.2 FPS and MCD 62.5% mIoU at 22.4 FPS on Jetson AGX Orin, both above a 10 Hz scan. Drop-in front end: outdoor odometry ATE RMSE −10.6% average, accepted-loop recall +8.7 pp on SemanticKITTI. Indoor sequences also evaluated. Code posted.
Why it matters
A Jetson-class LiDAR semantic FPS you can ask a perimeter / curb vendor to match, not a 4090 demo. Accelerated compute beats locked “AI cameras” (swap models vs a 36–84 month refresh).
Source
arXiv:2611.14025 - 7
AgenticConflictLens: query-conditioned conflict mining in long traffic video — 0.84 F1, inspects ~9–12% of hours
Key point
Boshra Khalili et al.: safety-critical events (conflicts, near-misses, unsafe turns, sudden braking, crosswalks) are rare in hours of fixed-camera traffic video. AgenticConflictLens takes a natural-language safety query, uses traffic memory plus conflict tools, coarse-to-fine search, then emits intervals, participant pairs, regions, trajectories, uncertainty, and rejection reasons. On 96 annotated hours (Bellevue and NYC DOT): Bellevue 0.84 F1 / 0.77 tIoU while inspecting 8.7% of video; NYC DOT 0.80 F1 / 0.73 tIoU at 11.9%.
Why it matters
City traffic VMS hours are a search problem. Ask vendors for query-conditioned F1 and fraction-of-video-read, not a 24/7 VLM bill.
Source
arXiv:2611.13540 - 8
EVAThink: evidence-sufficiency training so a video VLM answers or reasons under a fixed frame budget
Key point
Changti Wu et al.: under a fixed frame budget, some questions have enough sampled frames to answer directly; others need reasoning over sparse cues. EVAThink mines recoverable evidence-gap pairs, then two-stage RL with first-token regularization against mode collapse. Trained on 6.7K samples; improves the base VLM on video-understanding benches at 32- and 64-frame budgets. Code, model, and data posted.
Why it matters
On-prem video VLMs will be sold on frames-per-query. Ask whether the model skips thinking when the clip already has the plate/person, or always burns a long CoT.
Source
arXiv:2611.13907